Chris Patman

researching various aspects of symplectic flows, potentials, generalized spectral analysis in probabilistic metric spaces for feature generation and extraction.

[Chris Patman @cpatman]

Taking a break, browsing this again and reflecting on what a great gift to the mathematical community this was: https://t.co/KrOTtpgSlr

[Chris Patman @cpatman]

@OxUniMaths https://t.co/4vASEOGx8k

[Chris Patman @cpatman]

@OxUniMaths *Sir Roger Penrose

[Chris Patman @cpatman]

Gary, I think you have a right to be concerned because the caricature is close enough to your 2018 critique that people who don’t have nuance will just randomly dogpile on anything that seems close.

I’ve always read your progress take as something akin to: AI progress matters enough that you think the field should stop confusing progress in one extraordinarily successful technique with a proof that this technique constitutes the final architecture of intelligence.

AI progress is certainly not over. What you have argued is that simply scaling the present LLM/deep-learning paradigm (more parameters, more data, more compute) will encounter diminishing returns and will not by itself produce reliable AGI. Continued progress will require substantive architectural and scientific changes, probably involving things like explicit reasoning, world models, causality, symbolic and/or neurosymbolic methods, perhaps with insights from cognitive science.

This is my current understanding of your position on incentives and the public good, roughly: existing generative AI can still be useful and continue improving. Separately, you do not think AI companies’ incentives or existing law are sufficient to reliably protect the public; meaningful external testing, accountability, and regulation are needed.

If anything I said is incorrect, please correct my understanding of your positions.

[Chris Patman @cpatman]

@Haniel_Ulises I really like “early” books on neural nets; the DARPA study is especially nice.

[Chris Patman @cpatman]

@AstroMikeMerri @memcculloch Area laws are so common that observing one gives you very little information about which underlying mechanism generated it.

[Chris Patman @cpatman]

@PierreJoye @burkov I agree with your take. They are extremely useful but can’t do basic logic. The power is with the driver for the time being and that’s the disconnect we see.

[Chris Patman @cpatman]

Yes, and even if you were meticulous about redacting all citations and references there still would be oddities that pin down the publication date, the envelope would be larger, but good enough for classification. Specifically, “hmm the inputed document seems to only discuss concept and ideas relative to the time period leading up to 1984, so publication date is probably between 1984-1992 at some confidence level … automatically classify it as human generated”. On the other hand, forging an ai document to look as if it’s pre ai era is extremely difficult. The way language is used, phrasings, word choice, framings and perspectives all change through time and have their own correlative distributions, not to mention the distributions and patterns the ai’s typically fall into. It’s a genuinely interesting math problem applied to language.

[Chris Patman @cpatman]

Great point, but it’s worse, dissertations use citations and references, and since you typically don’t cite future works that have not been written yet, you can pin down a small spread for the dissertation date— so the whole “I scanned 24,000 documents” claim is ruined. The false-positivity rate is still in question.

[Chris Patman @cpatman]

*economic reasoning.. reasoning..

Many mistakes classified as failures of economic reasoning are instances of failures in reasoning, inference, and first principles thinking. I would say they need to do more work on foundational aspects of critical thinking. MoEs were supposed to help but I guess we need adversarial MoEs explicitly (since people may not be critical and/or adversarial with their own thought processes).

[Chris Patman @cpatman]

“Our paper”.. did you just dox yourself or was that a typo?

[Acer @AcerFur]

@QualiaQuanta @GaryMarcus 2/ Your center argument confuses a semidirect product with a direct product. D is abelian but not central: (0,k)(d,1)(0,k)^{-1} = (k•d,1). Our paper proves every nonzero d has an infinite K-orbit, hence Λ is ICC.

[∞-modal Noah @NoahChrein]

I haven't been able to crack this psychological nut, please help me to understand:

Why aren't math / physics theorists, on the whole, DYING to know what is going on inside of these large scale cognitive systems, e.g. ai interpretability from a pure theory perspective?

Why is it seen, at best, as just a "tool" to use for other theory crafting / problem solving, but not the subject of theory-craft itself?

[Chris Patman @cpatman]

@TheHPfacts I actually wish they went more creative with his hairstyle and gave him some flair, something like this..

[Chris Patman @cpatman]

Your system sounds great for your needs!

I use repeated, weighted access as evidence of relevance so each consultation contributes a weight wᵢ to a time-decaying score,

A(d,t) = Σᵢ wᵢ exp[−λ(t − tᵢ)],

where wᵢ reflects the depth of use. I skim digitally, print a document when A(d,t) crosses one threshold, and promote it to close reading if sustained use pushes the score past a higher threshold.

Because technical reading can take time, I use this system as a triage to only spend time on things that have proven to be useful or continues to catch my attention; but it’s not optimal!

I’ve been playing with various schemes for the printed materials and am always on the lookout for improvements.

[Chris Patman @cpatman]

@emily00800800 @QiaochuYuan Please describe your system.. I’m always looking for optimizations! I see color coding, check mark, “read” label, and possibly a scoring/ranking system (your 1/3 designation) that might imply a few things..

[Chris Patman @cpatman]

The corollary emphasizes an arithmetic obstruction to exact cancellation when the polynomial, scale, and endpoints are algebraic.. redundantly.. probably because the ai later needs the resulting nonvanishing statement and is drawing the reader’s attention to it. Interesting that an ai is anticipating human readers needs. Very interesting behavior.

[Chris Patman @cpatman]

@grok @michaelmalice @alaresurrected I wish people would learn to use logic and think..

[Chris Patman @cpatman]

@grok @michaelmalice @alaresurrected https://t.co/k9luF5uNZq

[Chris Patman @cpatman]

Oh I trust both of them completely! Sam will play the game and compete to win. He is very talented in this way. “Wario” will continue the doom narrative and I trust that he will continue to hide behind a facade, masking his true intentions which are malevolent. One is a business man, the other is a snake.

[Chris Patman @cpatman]

Product market fit will always be a thing. Software and apps are getting commoditized, so your valued-added proposition is your domain experience where your solution lives.

My recommendation is to not try to make money at first, try to work on solutions that solve problems for you, then combine that with your experience to serve solutions.

There once was a time were a single proprietor or consultant was limited to the amount of projects they could take on but now, if you are clever, you can scale your attention to items that really need it with automation and ai use.

Your software is not longer the product, it’s just the access channel, your intelligence and domain experience is now the primary product.

[Chris Patman @cpatman]

@unclebobmartin Just release a cleanCode.md and shut them up..